基于纹理分析的脂肪肝B超图像识别

目的 为B超诊断脂肪肝建立计算机辅助诊断手段。方法 通过分析正常肝和脂肪肝B超图像的纹理特征，包括灰度共生矩阵的角二阶矩、熵和反差分矩统计特征，组成特征矢量，再用k-平均聚类算法和自组织特征映射人工神经网络算法对特征矢量进行分类处理。结果 k-平均聚类算法对正常肝的识别率为63．6％，对脂肪肝的识别正确率达90．9％；自组织特征映射人工神经网络对正常肝的识别正确率达84.8％，对脂肪肝的识别正确率达90．9％。结论 本文中建立的方法能较肉眼更精确地反映正常肝和脂肪肝B超图像的特征，如果再结合医生的临床经验能大大提高脂肪肝的诊断准确性。

B-Scan Ultrasonic Image Recognition of Fatty Liver Based on Texture Analysis

Objective To provide a computer-aided method for diagnosis of fatty liver by B-scan ultrasonic imaging. Method Fatty liver referred to the infiltration of triglycerides and other fats into liver cells, which affected the texture of the liver tissue. In this paper, texture features including angular second moment, entropy and inverse differential moment were calculated from gray-level co-occurrence matrices of B-scan ultrasonic liver images. Feature vectors indicating two classes of the images were established with the three features. Then these vectors were classified using k-means clustering algorithm and self-organized feature mapping (SOFM) artificial neural network. Result The accuracy rates of k-means clustering algorithm were 63.6% for normal liver and 90.9% for fatty liver. The neural network algorithm showed accuracy rate of 84.8% for normal liver and 90.9% for fatty liver. Conclusion This technology shows the characteristics of B-scan images of both normal liver and fatty liver more accurately than eyes do. It greatly improves the diagnosing accuracy of fatty liver.